FRIDAY · SEPTEMBER 18, 2026 · ISSUE 13 · SINCE MONDAY No. 13
Daily News — 2026-09-18
1,607 papers indexed on arXiv·~10,000 packages released on PyPI·107 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
A tabular foundation model that learns causal graph structure as a byproduct of masked prediction, beating dedicated causal discovery methods without any task-specific training.
The central claim here is architectural: standard tabular foundation models organize prediction around a single designated target column, and that confinement limits what they can learn about the joint structure of a…
Deletion-only context compaction that preserves exact values by never rewriting anything — a principled trade-off with real costs worth understanding before deploying.
Context compaction in long agentic sessions is a real problem, and the usual answer — ask a model to summarize old turns — has a real flaw: summaries drop exact values. A file path, a precise error message, a constraint…
Papers
A tabular foundation model that learns causal graph structure as a byproduct of masked prediction, beating dedicated causal discovery methods without any task-specific training.
LimiX-2 is a tabular foundation model whose joint context-dependent pretraining also enables causal skeleton recovery from feature attention.
Sparse critic supervision at three positions per response fixes a structural flaw in PPO that dense token-level training was silently creating.
PPO critics in LLM training silently flatten value estimates; supervising only 3 sparse states per response (SP^3O) fixes it and improves policy quality.
A training-free confidence estimator that beats ten-sample self-consistency on 23 of 24 benchmarks by asking a model to consult its own graded history rather than re-read its current reasoning.
XConf estimates model confidence by retrieving graded past episodes on similar tasks and restating confidence informed by historical success rates.
A benchmark that forces agents to discover specifications from running software, exposing a sharp gap between observing behavior and faithfully reproducing it.
A benchmark that evaluates coding agents on features inferred from fully functional reference web apps, covering 4,063 tasks across 26 applications.
A credible coordination substrate for multi-agent research, with one honest gap: the causal comparison against a plain leaderboard hasn't been run.
Agora gives multi-agent research loops a shared Git DAG: 13 uncoordinated LM workers ran 12 days and closed 62% of the gap to a trained GPT-2.
Sorting value tokens by similarity before quantization, not rotating them, is what closes the dominant accuracy gap in low-bit video diffusion attention.
Training-free low-bit attention for video diffusion: value-token reordering plus a fused ExpCast-FP8 deliver 1.46-1.59x kernel speedup on Blackwell/Hopper.
9 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A real-time joint-control world model that runs cheaply and works, with an unusually clear-eyed account of why persistent state remains unsolved.
Zing-0.5 is a 5B world model with joint keyboard+text control running at 24 FPS for ~$0.009/stream-minute, weights released.
Tools & packages
Deletion-only context compaction that preserves exact values by never rewriting anything — a principled trade-off with real costs worth understanding before deploying.
A Claude Code plugin that scores every tool call and result in one request, dropping or truncating stale entries while keeping the rest verbatim.
A no-dependency toolkit that actually reads the undocumented local databases your AI coding tools have been quietly accumulating.
An open-source tool that extracts chat histories from AI coding assistants including Claude Code, Cursor, Windsurf, Aider, and Cline/Roo Code.
A tightly scoped, version-pinned automation bridge between JSON edit plans and native CapCut drafts — honest about every edge it hasn't crossed.
A headless interface for Jianying that supports draft editing, export, and a standalone agent skill.
A working App Store request-tampering guide that illustrates client-side purchase validation failures — and describes conduct that violates Apple's and OpenAI's terms of service.
A security analysis of the ChatGPT iOS subscription and entitlement flow, written as an internal draft under coordinated disclosure.
A clean argument that pre-computed, object-centric JSON beats raw observation for model-driven game control — worth reading for the harness design alone.
An agent that plays Super Mario Bros. using structured emulator state.
Runtime refusal ablation that keeps ternary weights bit-identical — a technically honest solution to a real quantization constraint, with unusually candid evaluation caveats.
Runtime behavioral ablation for a 27B compressed LLM, applied without weight modification or re-quantization.
7 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A versioned, permission-scoped skills store for agents that treats prompts like code — with real revision control and no shared provider accounts.
Self-hosted skills library for AI agents with versioned MCP tools, scoped clients, and optional Jev recommendations.
A local-first Mac workspace that keeps AI edits, source documents, and project memory in one place — most interesting for its research wiki and diff-review design.
Local-first desktop workspace that ties AI chat to source-linked knowledge and a built-in task tracker.
A versioned, self-correcting semantic layer that measurably reduces the repeated data-exploration tax on LLM agents working over real databases.
EvoOntology adds a self-evolving ontology layer to data agents running in Claude Code and Codex.